Ethical implications of machine learning

Ethics of technology in daily life

To discuss machine learning ethically, weigh its benefits against possible harm. Ask who is affected and whether the safeguards are sufficient. Even a model whose predictions are mostly accurate can treat some people unfairly.

### Ethical responsibilities

Imagine Cedar Housing, an invented housing service that uses machine learning to prioritize maintenance requests. Spotting urgent faults faster could protect residents. Incorrect decisions, though, could leave some households waiting for essential repairs.

| Ethical issue | What must be considered in this scenario? |
|---|---|
| **Accountability** | Who must investigate an incorrect priority decision and put it right? The service remains responsible to residents even when it uses a model. |
| **Algorithmic fairness** | Are requests handled fairly across different groups of residents, or does one group consistently lose out? |
| **Bias** | Do the data or the way decisions are made systematically favour some residents over others? |
| **Consent** | Do residents understand and meaningfully agree to how their information is used? Agreement to report a fault does not automatically cover unrelated uses. |
| **Privacy** | Is personal information collected and used only where needed, with access restricted appropriately? |
| **Security** | Is information protected against unauthorized access or alteration that could expose residents or change priorities? |
| **Transparency** | Can residents understand the system's role and the main reasons for a decision? Do they know how to challenge it? |
| **Environmental impact** | Do the benefits justify the energy and resources used to train and operate the model? |
| **Societal impact** | Does automation improve access to repairs, or leave residents dependent on a service they cannot effectively question? |

These issues overlap. Different settings may raise other ethical concerns too. Collecting more personal information, for example, might help detect unfair outcomes while increasing privacy risks. A reasonable response must weigh both effects; neither principle automatically settles the matter.

### Challenges caused by training-data bias

Some users may be underrepresented in training data. The records may also reflect earlier unfair decisions or contain inconsistent labels. A model learns patterns from those records, but it does not independently judge whether those patterns are ethically acceptable.

Suppose most of Cedar Housing's historical reports came through a daytime telephone service. Residents who could only report faults in the evening might then be poorly represented. The figure traces a possible chain: **unrepresentative training data** shape **learned patterns**, which influence **priority decisions** and lead to **unequal waiting times**. When those decisions become **new training records**, a **feedback loop** can reinforce the original disadvantage.
{{0779262a6ce2be1f31488709d48dbd0934ca25c8}}

Consider this small, invented evaluation. The correct priority for each request has been checked independently, allowing us to see why overall accuracy is not enough.

| Reporting group | Requests checked | Correct priority decisions | Incorrect priority decisions |
|---|---:|---:|---:|
| Daytime reporters | 4 | 4 | 0 |
| Evening reporters | 4 | 2 | 2 |
| Total | 8 | 6 | 2 |

The model gets 6 out of 8 requests right: 75% overall. It is correct for every daytime request, yet only half the evening requests. This small sample points to a problem worth investigating. It neither proves the cause nor establishes how the model performs across all residents.

Adding more records will not necessarily fix the problem. If they come from the same unrepresentative source, they may preserve it. Appropriate responses include checking representation and reviewing labels, then evaluating outcomes for affected groups and monitoring decisions after deployment. Human review and a route to challenge decisions can reduce harm. To correct errors, however, reviewers need enough information and the authority to act.

There is a genuine tension here: removing useful information may reduce predictive accuracy, while keeping it may reproduce unfair historical patterns. Judge the consequences of errors for different people, rather than relying only on the model's overall score.

### Machine learning in online communication

Willow Forum is an invented discussion platform that uses machine learning to recommend posts and flag harmful messages. These tools can support communication or cause harm; much depends on how they are designed and used.

| Concern | Potential benefit | Ethical risk and possible response |
|---|---|---|
| Misinformation | Automated checks can quickly flag potentially misleading material. | Incorrect flags can suppress reliable information, while missed material can spread widely. Review disputed decisions and explain the limits of automated checks. |
| Bias | Moderation can apply shared rules across many posts. | Uneven training data may lead to some dialects or communities being flagged more often. Evaluate outcomes across affected groups. |
| Online harassment | Early detection can reduce exposure to threatening messages. | Misunderstood context may leave victims unprotected or lead to legitimate speech being removed. Provide reporting and review processes, with a route to appeal. |
| Anonymity | People can discuss sensitive experiences without publicly revealing their identity. | Anonymity can also make abusive behaviour harder to address. Balance protection for vulnerable speakers with proportionate accountability. |
| Privacy | Detecting harmful patterns can help protect users. | Analysing messages or personal behaviour can become intrusive surveillance. Limit what information is collected, who can access it and how it is reused. |

Other concerns may matter too. No single setting eliminates every risk: stricter moderation may reduce harassment while removing more legitimate speech. Ethical deployment requires continuing evaluation and user involvement, with clear responsibility for harmful outcomes.

**Key terms:** accountability, algorithmic fairness, bias, consent, environmental impact, privacy, security, societal impact, transparency, training data, misinformation, online harassment, anonymity.

A left-to-right causal diagram labelled unrepresentative training data, learned patterns, priority decisions and unequal waiting times. A return path through new training records to unrepresentative training data is labelled feedback loop. Show how biased records can produce and reinforce unequal outcomes.

As technology becomes a larger part of daily life, we need to weigh convenience and new capabilities against effects on society, individual rights, privacy and equity. Many people may benefit from a technology while others bear a disproportionate share of its costs.

### Why ethical guidelines need continual reassessment

Ethical guidelines can become inadequate as a technology's capabilities, scale or uses change. A system might begin by analysing occasional user requests, then shift to monitoring behaviour continuously. The original guidance may not cover the privacy risks this creates.

Unexpected harm revealed by evidence also calls for reassessment. So does a technology reaching new groups, or becoming so widely adopted that opting out is difficult. Consent given at the start does not settle every ethical question about later uses.

The figure shows a repeating process: **technology or use changes**, followed by **identify affected people**, **assess benefits and harms**, **revise ethical guidelines and safeguards**, and **monitor outcomes**. Monitoring may reveal further changes. The process therefore repeats rather than stopping at approval.
{{86d68e1934e96cf39ce5ea5a52b36fb7e3667df5}}

Reassessment should include those who benefit as well as those who bear the risks or cannot gain access. Developers and service providers may understand what the technology can do. Affected users, though, can reveal consequences that technical testing alone misses.

### Emerging technologies and their implications

| Technology | Potential benefits for society | Ethical concerns affecting rights, privacy or equity |
|---|---|---|
| **Quantum computing** | Could offer new ways to approach difficult computational problems. | Future capabilities could weaken some current protections for private information. Expensive infrastructure could leave access and influence concentrated among well-resourced organizations. |
| **Augmented reality** | Overlaying information on the physical environment could help with practical tasks and improve access to guidance. | Cameras and sensors may record bystanders who have not agreed to recording. Some users may be excluded by cost or accessibility barriers. |
| **Virtual reality** | Immersive environments could give people more opportunities for training and shared experiences. | Detailed records of movement and interaction can reveal personal behaviour. Users can experience harassment, while equipment costs or accessibility limitations can restrict participation. |
| **Pervasive use of AI** | Assistance built into everyday services can improve convenience and make some tasks more accessible. | Continuous data collection can reduce privacy; automated decisions can limit autonomy or reproduce discrimination. People may be disadvantaged if they lack suitable access or skills. |

None of these outcomes is guaranteed. Their significance depends on system design and control, as well as the setting in which the systems are used. Future quantum-computing risks, for example, justify reviewing information protection. They do not justify claiming that every present-day system is already insecure.

### Weighing integration against meaningful choice

Suppose a fictional training centre, Harbour Skills, introduces augmented reality instructions and virtual reality practice. Learners may gain opportunities to practise tasks safely. Taking part, however, could require expensive equipment or the collection of detailed behavioural information.

Providing equipment could reduce financial exclusion without resolving every accessibility barrier. A non-immersive alternative could preserve choice, though maintaining alternatives has costs. Collecting less data would protect privacy but might leave less detail available to personalize instruction.

AI assistance presents a similar tension: it may increase independence for some users while making others dependent on automated recommendations. Giving everyone the same interface does not necessarily make a service fair. People have different resources and accessibility needs, and differ in their ability to challenge decisions.

The ethical judgement should therefore be conditional. Integration is more defensible when evidence supports the benefits and harms are proportionately controlled. Access must also be equitable, with people retaining meaningful choices. These conditions need review as the technology and its effects change.

**Key terms:** ethical guidelines, emerging technologies, quantum computing, augmented reality, virtual reality, AI, society, individual rights, privacy, equity.

A circular reassessment flowchart ordered technology or use changes, identify affected people, assess benefits and harms, revise ethical guidelines and safeguards, and monitor outcomes. Connect monitor outcomes back to technology or use changes to emphasize continuing ethical review.

Machine learning approaches

Identify two ethical risks associated with using this model.

Suggest two measures Marden Cycles could take to reduce the risk of unfair shortlisting.

Identify two ethical risks arising from the use of machine learning on Kestrel Exchange.

Suggest two safeguards that would allow users to challenge or correct harmful decisions made by these models.

Identify two ethical concerns introduced or increased by the proposed update.

Suggest two revisions to Orlin Visitor Service’s ethical guidelines that would address the changed capabilities.

Computer Science

IB Computer Science focuses on computational thinking and programming skills. Students learn about system fundamentals, network architecture, and software development, applying their knowledge to solve real-world problems through practical projects.

A4.4 Ethical considerations

Question 1

[2]

[2]

Question 2

[2]

[2]

Question 3

[2]

[2]